The Emergence of Cultural Attractors: How Dynamic Populations of Learners Achieve Collective Cognitive Alignment

Falandays, James B., and Paul E. Smaldino. 2021. “The Emergence of Cultural Attractors: How Dynamic Populations of Learners Achieve Collective Cognitive Alignment.” PsyArXiv. July 28. doi:10.31234/osf.io/v5fsr

Abstract: Cultural attractor landscapes describe the time-evolution of cultural variants over successive transmission events. Because these landscapes are statistical patterns that emerge from the interactions of dynamic learners in dynamic populations/environments, stable landscapes cannot be taken for granted. However, they are often modeled as such, and little is known about how attractors form, change, and/or stabilize. We present a model of cultural attractor dynamics, which adapts a model of unsupervised category learning in individuals to a multi-agent setting, wherein learners provide the training input to each other. Agents in our populations spontaneously align their cognitive category structures, producing emergent cultural attractor points. We highlight three interesting behaviors exhibited by our model: (1) noise enhances stability of cultural category structures; (2) short “critical” periods of learning early in life enhance stability; and (3) larger populations produce more stable but less complex attractor landscapes, and cliquish network structure can mitigate the latter.

---

All human groups possess group-specific behavioral repertoires involving cultural variants—things such as tools, linguistic behavior, social norms, religious beliefs, and artistic styles. As cultural variants are observed and copied, they are liable to change over time due to the accumulation of errors in transmission. However, even in the absence of strong selection for specific outcomes, cultural variants may nevertheless converge over successive transmission events toward culture-specific “attractor” points (Sperber, 1996). This effect can be attributed to the fact that individuals within a cultural group tend to have similar cognitive biases, such that they tend to perceive, remember, and reproduce information in consistent ways (Heyes, 2018). Without this "cognitive alignment," cultural transmission would be far less reliable, and the potential for cumulative cultural evolution would be limited. But how does cognitive alignment first emerge in dynamic populations? Current models of cultural evolution take cognitive alignment as given. However, this assumption may not always be justified, since many aspects of culture depend on cognitive biases that are themselves socially learned (Heyes, 2018; Karmiloff-Smith, 1994). As each new generation learns through exposure to the cultural products of the previous generation, they may acquire different cognitive biases than their teachers, in turn resulting in a new set of cultural products in the next generation. An unstable feedback loop of this kind could disrupt the accumulation of cultural knowledge in a population, as later generations may no longer perceive, remember, and reproduce information in ways consistent with their ancestors. Furthermore, populations are not static: new individuals are born or enter the population from elsewhere, while others die or leave for new lands. Within such shifting populations, cognitive alignment needs to be actively and continuously maintained in order for cultural knowledge to be successfully preserved across generations. In this paper, we develop an agent-based model of the emergence and maintenance of cognitive alignment in dynamic populations, where individuals act as both teachers and learners to each other, and refine their cognitive biases over the course of interaction. Our initial explorations with this model suggest that achieving and maintaining cognitive alignment may depend upon a finely tuned balance of factors at the levels of cognition, development, and demographic structure. We highlight three interesting and potentially counter-intuitive behaviors exhibited by our model that are not accounted for in other models of cultural evolution: First, we find that some noise is beneficial to stabilizing cognitive alignment. Second, we find that long learning times may destabilize and limit the complexity of cultural repertoires, while critical or sensitive periods of learning enhance stability. Third, we find that larger populations develop less complex, but more stable patterns of alignment, and that this effect can be moderated by network structure. These results suggest that additional complexity may be needed in models of cultural evolution to adequately understand how human-level culture can get off the ground and develop. We conclude by highlighting several ways that our model may be extended to complement 40 existing models of cultural evolution and gene-culture co-evolution.